From a photo to a talking clip

A common job: take one photo, put the person in a new place, and have them say something. Doing it in two steps gives you a look at the frame before you pay for the clip.

The commands use jq to pick fields out of the answers. RUN makes this run's idempotency keys its own, so you can run the guide more than once:

RUN=$(date +%s)

1. Upload the photoCopy Anchor Link

PHOTO_ID=$(curl -s https://api.nocamclub.com/v1/uploads \
  -H "Authorization: Bearer $NOCAM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "url": "https://docs.nocamclub.com/examples/mascot.jpg" }' | jq -r .id)

Use your own photo's address instead. The id works for 48 hours.

2. Make the frameCopy Anchor Link

Vertical, since that's what the clip will be. Say what to keep.

FRAME_URL=$(curl -s https://api.nocamclub.com/v1/generations \
  -H "Authorization: Bearer $NOCAM_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Prefer: wait=60" \
  -H "Idempotency-Key: rooftop-frame-$RUN" \
  -d '{
    "model": "nano-banana-2",
    "input": {
      "prompt": "Keep the face and hair of img1. The same person on a city rooftop at sunset, facing the camera, waist up. Phone photo.",
      "images": ["'"$PHOTO_ID"'"],
      "resolution": "2K",
      "aspect_ratio": "9:16"
    }
  }' | jq -r '.outputs[0].url')

Look at the result. If it isn't right, change the prompt and make another: a photo costs little and a clip costs more.

3. Make it speakCopy Anchor Link

A result is a public link, so it goes straight to the next model.

Muse always makes 10 seconds with sound, and its shape follows the photo. About 2 words a second fits the length.

CLIP_ID=$(curl -s https://api.nocamclub.com/v1/generations \
  -H "Authorization: Bearer $NOCAM_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: rooftop-clip-$RUN" \
  -d '{
    "model": "muse-video",
    "input": {
      "prompt": "img1 looks straight into the camera, smiles and says clearly: \"I made this whole video from one selfie.\" The camera does not move.",
      "images": ["'"$FRAME_URL"'"]
    }
  }' | jq -r .id)

Then poll GET /v1/generations/$CLIP_ID until status is succeeded or failed, usually a minute or two.

Why two stepsCopy Anchor Link

• You check the expensive part's input. A bad frame costs a few credits. A bad clip costs more, and takes longer to find out.
• You choose per step. Use the model that's best at photos, then the one that's best at clips.
• Failures stay small. If the clip fails, the credits come back and the frame is still there.

Keep it working in productionCopy Anchor Link

• Send a new Idempotency-Key per clip, and reuse it for retries of that clip.
• Read error.code: on content_refused, tell the person. On provider_unavailable, try again in a few minutes.
• Download the result: links stop after 7 days.